Application of Neural Networks and Other Learning Technologies in Process Engineering

Transcription

1 Application of Neural Networks and Other Learning Technologies in Process Engineering

2 This page is intentionally left blank

3 Published by Imperial College Press 57 Shelton Street Covent Garden London WC2H 9HE Distributed by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Farrer Road, Singapore USA office: Suite IB, 1060 Main Street, River Edge, NJ UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library. APPLICATION OF NEURAL NETWORKS AND OTHER LEARNING TECHNOLOGIES IN PROCESS ENGINEERING Copyright 2001 by Imperial College Press All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher. For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher. ISBN Printed in Singapore.

4 To my parents: Professor M. Ishaque and R. Akhter My wife: Nasreen And my children: Sumayya, Maria, Hamza and Usama I.M. Mujtaba To my parents: Hussain Mohamed and Khairun Haider My wife: Fakhriani Hj. Yusof And my children: Nor Daleela, Ahmad Nasruddin, Ahmad Zubair, Nor Sakeenah and Nor Ameenah M.A. Hussain

5 This page is intentionally left blank

6 Foreword This book is a follow-up of the IChemE CAPESG workshop on "The Application of Neural Networks and Other Learning Technologies in Process Engineering" held on the 12th May 1999 at Imperial College, London. The interest showed by the participants especially those from the industries in these emerging technologies has inspired us to come up with this book. This is not only the proceedings of the workshop but an expanded and revised versions of the talks presented at the workshop as well as invited papers from other well known international researchers in this area. Hence in short, this book contains contributions in the field of neural networks and learning technologies from experts in different parts of the globe. In summary the papers are arranged in this book in parts based on certain topic-related sequences. Part I (Papers 1 to 5) relates to the use of neural networks for identification and modelling purposes as well as some aspects of neural network training. Part II (Papers 6 to 8) discusses the utilisation of neural networks in hybrid schemes for modelling and control purposes. Part HI (Papers 9 to 11) relates to the use of this technology for estimation and control of various chemical processes. Part IV (Papers 12 and 13) involves their usage in new and learning technologies strategies in chemical process systems while Part V (Papers 14 and 15) discusses the use of this technology in experimental and industrial applications. Part I: Modelling and Identification The first paper by Aldrich and Slater starts with the discussion on the use of neural networks for modelling of liquid-liquid extraction column and prediction of equilibrium data and kinetic coefficients. In the paper they show examples of the use of neural networks for dispersed phase holdup and drop size prediction in extraction columns and rotating disc contactors. They also demonstrated the modelling of extraction in a vortex ring batch cell as Vll

7 Vlll Neural Networks in Process Engineering well as the performance monitoring of extraction in an industrial column using neural network methodology. The next paper by Bomberger et al. is about utilising radial basis function (RBF) networks for the identification of a multivariable coplymerisation reaction in a continuous stirred tank reactor. The k-means clustering and stepwise regression analysis methods are used for the process of RBF modelling. The minimum model order is determined using the method of false nearest neighbours. The simulation is also performed utilising conditions similar to the actual plant to assess its practical approach. The third paper by Eikens et al. demonstrates the use of unsupervised neural networks in the form of self-organising maps for process identification in a yeast fermentation system. The network was found to predict accurately the different physiological states in the fermentation process. The forth paper by Kershenbaum and Magni is about the use of nonlinear techniques to determine the proper centre locations in radial basis function networks. The training approach is performed through the Bayesian method and done for a simulated continuous stirred tank reactor system and a kin robot arm utilising Gaussian and thin plate spline networks. This approach is found to improve the performance of the networks over that of the traditional unsupervised methods. The fifth paper by Scheffer and Maciel Filho involves the use of a recurrent neural network for nonlinear identification of a fed-batch penicillin process. In this work the neural network is trained by a multiple stream extended Kalman filter methodology. This approach allows the processs to be identified in real time which is a useful tool for calculation of the optimal feeding strategy in real time. Part II: Hybrid Schemes Paper 6 by Eikens et al. utilises first principles parametric models with neural networks in a hybrid strategy to identify a fed-batch fermentation process. Different types of neural networks were integrated into the hybrid model structure in the simulation work for multi-step ahead predictions and these results were compared with utilising the traditional neural network approach.

8 Foreword IX Paper 7 by Greaves et al. discusses the use of neural networks in hybrid strategies for optimal control purposes. In this paper a hybrid model for an actual pilot batch distillation column is developed where the neural network is used to predict the plant-model mismatch of the system. With this hybrid model, a general optimisation framework is developed to find optimal reflux ratio policies which then minimises the batch time for a given separation task. Paper 8 by Meleiro et al. discusses the use of the hierarchical neural fuzzy models in the simulation of an industrial plant. The models here consist of a set of radial basis function networks formulated as simplified fuzzy systems connected in cascade. This hybrid model approach is then applied for the modelling of a multi-input multi-output complex biotechnological process for ethyl alcohol production with long range prediction capabilities. Part III: Estimation and Control Paper 9 by Hussain involves the control of a continuous fermentation process, using the internal-model control strategy, wherein the neural network inverse model act as the controller in the closed loop system. The simulation for the control of the biomass concentration was performed for both set point tracking and disturbance rejection cases. The offsets obtained in these cases were eliminated by the use of an adaptive online control scheme, wherein the adaptation of the forward and inverse models was carried out. Paper 10 by Aziz et al. demonstrates the use of neural networks for estimating the heat released in an exothermic batch reactor system. This estimation was then used in a generic model control scheme for controlling the reactor temperature by manipulating the jacket temperature. The set point tracking of the reactor temperature followed an optimum profile generated by the formulation of the reactor's optimal operation in the offline mode. Comparisons with the conventional dual mode strategy were also shown in this work. The next paper (paper 11) by Zhang and Morris utilises the bootstrap aggregated stacked neural networks approach to nonlinear empirical modelling. This method is effective in building models from a

9 X Neural Networks in Process Engineering limited data set. In their study, the robust neural network was utilised for inferential estimation of polymer in a batch polymerisation reactor. The estimation of the amount of reactor fouling during the early stage of the batch process was also done as well as the optimal control of the batch polymerisation process. Neural network models are used to provide inferential estimation of polymer quality as well as to predict the trajectory of polymer quality variables from the batch recipe and control profile, which provide appropriate control actions for the polymerisation process. Part IV: New Learning Technologies Paper 12 by Wilson and Martinez utilises the reinforcement learning method for optimisation and control of a semi-batch reactor process. They utilise the notion of the performance of the value function to achieve the target. For batch-to-batch learning and control, the value function is represented by wire fitting methods incorporating neural networks methodology. The next paper (paper 13) by Wang demonstrates the use of the emerging data mining and knowledge discovery technology in analysing large volumes of data in a meaningful way. One case study involves utilising data from a refinery separation process to help operators in analysing the operational states of the process. The second case study involves utilising wavelet analysis for identifying feature extraction and operational states in a fluid catalytic cracking process while another study on a methyl tertiary butyl ether plant illustrates the clustering approach in identifying the operational states of the process. Part V: Experimental and Industrial Applications The paper 14 by Cabassud and Le Lann involves neural networks in three experimental applications. The first one involves utilising neural networks in an inverse model method to control a semi-batch chemical reactor pilot plant with time varying operating conditions. Various neural network designs were investigated in this study. The second study involves using neural networks in a mutivariable controller for controlling a liquidliquid extraction column. The control strategy was done based on the

10 Foreword XI inverse modelling approach. The results obtained showed improvement with regard to previous studies of using the conventional adaptive control method. The third study involves using neural networks to measure and control a low-pressure chemical vapour deposition reactor. A hybrid neural network model was developed to compute the deposition rate profile along the reactor. A mutivariable controller using inverse dynamic methodology was also developed to compute the local set points of the PID controllers. The last paper 15 by Puigjaner discusses the use of neural networks in evolutionary optimization of a nonlinear, time-dependent process in combination with genetic algorithms. Neural network is used off-line to update real plant representation and for multilevel decision making online as well as in real time optimisation process. Results from various real industrial applications are reported and discussed in the paper.

11 Acknowledgements Alhamdulillah- All praise to almighty Allah who made it possible for us to complete this book. We thank IChemE CAPE subject group to give I. Mujtaba the opportunity to organize the symposium on "The Application of Neural Networks and Other Learning Technologies in Process Engineering" on 12 May The main inspiration for compiling such a book came from this symposium. Special thanks go to all the speakers of the symposium who accepted our invitation to contribute in this book. This book includes contributions from Europe, North America, South America, Africa and Asia. We are sincerely grateful to all the contributors who had sacrificed their valuable time to prepare the manuscripts. We would like to thank the reveiwers who made relentless efforts to review each manuscript carefully and to make useful comments. We gratefully acknowledge the UK Royal Society financial support to: (i) M.A. Hussain in 1999 for his visit to Bradford University when the initial planning to compile such a book was made; (ii) I. Mujtaba to cover the expenses in Malaysia during the final editing stage of this book. Finally, we thank to the publisher for publishing this book and sincerely acknowledge their support and help. Xll

12 Contents Foreword Acknowledgements Part I: Modelling and Identification 1. Simulation of Liquid-Liquid Extraction Data with Artificial Neural Networks C. Aldrich and M.J. Slater 3 2. RBFN Identification of an Industrial Polymerization Reactor Model J.D. Bamberger, D.E. Seborg, B.A. Ogunnaike Process Identification with Self-Organizing Networks B. Eilcens, M.N. Karim and L. Simon Training Radial Basis Function Networks for Process Identification with an Emphasis on the Bayesian Evidence Approach L.S. Kershenbaum and A.R. Magni Process Identification of a Fed-Batch Penicillin Production Process Training with the Extended Kalman Filter R. Scheffer, R.M. Filho 99 vii xii Xlll

13 XIV Neural Networks in Process Engineering Part II: Hybrid Schemes 6. Combining Neural Networks and First Principle Models for Bioprocess Modeling B. Eikens, M.N. Karim and L. Simon Neural Networks in a Hybrid Scheme for Optimisation of Dynamic Processes: Application to Batch Distillation M.A. Greaves, I.M. Mujtaba and M.A. Hussain Hierarchical Neural Fuzzy Models as a Tool for Process Identification: A Bioprocess Application L.A.C. Meleiro, R.M. Filho, R.J.G.B. Campello and W.C. Amaral 173 Part III: Estimation and Control 9. Adaptive Inverse Model Control of a Continuous Fermentation Process Using Neural Networks M.A. Hussain Set Point Tracking in Batch Reactors: Use of PID and Generic Model Control with Neural Network Techniques N. Aziz, I.M. Mujtaba and M.A. Hussain Inferential Estimation and Optimal Control of a Batch Polymerisation Reactor Using Stacked Neural Networks J. Zhang and A.J. Morris 243 Part IV: New Learning Technologies 12. Reinforcement Learning in Batch Processes J.A. Wilson and EC. Martinez 269

14 Contents xv 13. Knowledge Discovery through Mining Process Operational Data X.Z. Wang 287 Part V: Experimental and Industrial Applications 14. Use of Neural Networks for Process Control. Experimental Applications M. Cabassud, M.V. Le Lann Intelligent Modeling and Optimization of Process Operations Using Neural Networks and Genetic Algorithms: Recent Advances and Industrial Validation L. Puigjaner 371